A multiple stream architecture for the recognition of signs in Brazilian sign language in the context of health

Deaf people communicate naturally through sign languages and often face barriers to communicating with hearing people and accessing information in written languages. These communication difficulties are aggravated in the health domain, especially in a hospital emergency, when human sign language interpreters are unavailable. This paper proposes a solution for automatically recognizing signs in Brazilian Sign Language (Libras) in the health context to reduce this problem. The idea is that the system could assist in the communication between a Deaf patient and his doctor in the future. Our solution involves a multiple-stream architecture that combines convolutional and recurrent neural networks, dealing with sign languages' visual phonemes individual and specialized ways. The first stream uses the optical flow as input for capturing information about the "movement" of the sign; the second stream extracts kinematic and postural features, including "handshapes" and "facial expressions"; and the third stream process the raw RGB images to address additional attributes about the sign not captured in the previous streams. Thus, we can process more spatiotemporal features that discriminate the classes during the training stage. The computational results show that the solution can recognize signs in Libras in the health context, with an average accuracy, precision, recall, and f1-score of 99.80%, 99.81%, 99.80%, and 99.80%, respectively. Our system also performed better than other works in the literature, obtaining an average accuracy of 100% in an Argentine Sign Language (LSA) dataset, which is usually used for comparison purposes.